---
language:
  - hy
license: mit
base_model: Edmon02/TTS_NB_2
tags:
  - speecht5
  - onnx
  - text-to-speech
  - armenian
  - hy-am
  - tts
pipeline_tag: text-to-speech
library_name: transformers
inference: true
model_name: Armenian SpeechT5 ONNX
---

# Armenian SpeechT5 — ONNX export (`TTS_NB_ONNX`)

ONNX export of the SpeechT5 encoder/decoder for **faster or edge inference** (ONNXRuntime, mobile, C++). Pair with HiFi-GAN vocoder separately.

## Files

| File | Role |
|------|------|
| `encoder_model.onnx` | Text → hidden states |
| `decoder_model.onnx` | Autoregressive mel decoder |
| `decoder_with_past_model.onnx` | Decoder with KV cache |
| `decoder_postnet_and_vocoder.onnx` | Postnet (vocoder may still be separate) |
| `spm_char.model` | SentencePiece tokenizer |
| `config.json` / `preprocessor_config.json` | Model config |

## When to use

- ONNXRuntime deployment
- Environments without full PyTorch stack
- Latency-sensitive inference pipelines

For PyTorch + Hugging Face, prefer [Edmon02/speecht5_finetuned_voxpopuli_hy](https://huggingface.co/Edmon02/speecht5_finetuned_voxpopuli_hy) or [Edmon02/TTS_NB_2](https://huggingface.co/Edmon02/TTS_NB_2).

## Source checkpoint

Exported from [Edmon02/TTS_NB_2](https://huggingface.co/Edmon02/TTS_NB_2) (verify export date in commit history).

## Limitations

- ONNX export may not cover full HF `generate_speech` API — validate your runtime graph
- Speaker embeddings must be supplied consistently with training
- Re-export when `TTS_NB_2` architecture or opset changes

## License

MIT